def collect_leakage(device,
base_folder='./models/BAM/',
specific="bowling_alley",
seed=0,
module="layer3",
experiment="sgd_finetuned",
ratios=["0.0","0.1","0.2","0.3","0.4","0.5","0.6","0.7","0.8","0.9", "1.0"],
adv=False,
baseline=False,
epoch=None,
multiple=True,
force=False,
dataset='bam',
args=None):
results = {}
if dataset == 'bam':
_, testloader = dataload.get_data_loader_SceneBAM(seed=seed,ratio=float(0.5), specific=specific)
elif dataset != 'coco':
_, testloader = dataload.get_data_loader_idenProf('idenprof',train_shuffle=True,
train_batch_size=64,
test_batch_size=64,
exclusive=True)
for ratio in ratios:
model, net, net_forward, activation_probe = load_models(
device,
base_folder=base_folder,
specific=specific,
seed=seed,
module=module,
experiment=experiment,
ratio=ratio,
adv=adv,
baseline=baseline,
epoch=epoch,
post=True,
multiple=multiple,
leakage=True,
force=force,
dataset=dataset,
args=args
)
model.eval()
net.eval()
if dataset == 'coco':
tmp_args = copy.deepcopy(args)
tmp_args.ratio = ratio
tmp_args.gender_balanced = True
if int(ratio) > 0:
tmp_args.balanced = True
_, testloader = coco_dataload.get_data_loader_coco(
tmp_args
)
results[ratio],_ = utils.net2vec_accuracy(
testloader,
net_forward,
device,
train_labels=[-2,-1]
)
return results
specific=args.specific,
train_batch_size=args.train_bs)
_, testloader = dataload.get_data_loader_SceneBAM(
seed=args.seed,
ratio=float("0.5"),
specific=args.specific,
test_batch_size=args.test_bs)
elif args.dataset == 'idenprof':
trainloader, testloader = dataload.get_data_loader_idenProf(
'idenprof',
train_shuffle=True,
train_batch_size=args.train_bs,
test_batch_size=args.test_bs,
exclusive=True)
elif args.dataset == 'coco':
trainloader, testloader = coco_dataload.get_data_loader_coco(args)
# restart seed since we loop through dataloader to get mean/mu statistics
set_seed(args.seed)
device = torch.device('cuda:' + str(args.device))
assert (args.nonlinear and args.adv) or not args.nonlinear
args.experiment = args.experiment1
if args.finetuned:
args.experiment += "_finetuned"
# setup model directories
if args.dataset == 'bam':
if not isinstance(args.specific, str):
folder_name = '.'.join(sorted(args.specific))
else:
folder_name = args.specific
dir_parts = [
def collect_tcav(
device,
base_folder='./models/BAM/',
specific="bowling_alley",
seed=0,
module="layer3",
experiment="sgd_finetuned",
ratios=["0.0","0.1","0.2","0.3","0.4","0.5","0.6","0.7","0.8","0.9", "1.0"],
adv=False,
baseline=False,
epoch=None,
post=False,
multiple=True,
force=False,
dataset='bam',
args=None):
if dataset == 'bam':
_, testloader = dataload.get_data_loader_SceneBAM(seed=seed,ratio=float(0.5), specific=specific)
elif dataset != 'coco':
_, testloader = dataload.get_data_loader_idenProf('idenprof',train_shuffle=True,
train_batch_size=64,
test_batch_size=64,
exclusive=True)
model_type = "Baseline" if baseline else (
"Adversarial" if adv else "Debias"
)
df = {}
values = []
r = []
m = []
s = []
num_classes = 10
if dataset == 'coco':
num_classes = 79
for ratio in ratios:
print("loading model")
model, net, net_forward, activation_probe = load_models(
device,
base_folder=base_folder,
specific=specific,
seed=seed,
module=module,
experiment=experiment,
ratio=ratio,
adv=adv,
baseline=baseline,
epoch=epoch,
post=post,
multiple=multiple,
leakage=False,
force=force,
tcav=True,
dataset=dataset,
args=args
)
print("did load?")
if dataset == 'coco':
assert args is not None
tmp_args = copy.deepcopy(args)
tmp_args.ratio = 1
tmp_args.balanced = True
tmp_args.gender_balanced = True
tmp_args.train_bs = 64
tmp_args.test_bs = 32
_, testloader = coco_dataload.get_data_loader_coco(
tmp_args
)
print(net)
collect, labels = analysis.compute_tcavs(
testloader,
model,
activation_probe,
net.weight[-2]-net.weight[-1], #bias vector
device,
object_detect= (args.dataset == 'coco'),
objects=79
)
for s_idx in range(num_classes):
current = collect[labels==s_idx]
values.append(current)
r += [ratio] * current.shape[0]
m += [model_type] * current.shape[0]
s += [s_idx] * current.shape[0]
values = np.concatenate(
values
)
df['values'] = values
df['ratio'] = r
df['model'] = m
df['class'] = s
# return a Dataframe, then we can join together data from different experiments...
return pd.DataFrame(df)
def collect_accuracies(
device,
base_folder='./models/BAM/',
specific="bowling_alley",
seeds=[0,1,2,3,4],
module="layer3",
experiment="sgd_finetuned",
ratios=["0.0","0.1","0.2","0.3","0.4","0.5","0.6","0.7","0.8","0.9", "1.0"],
adv=False,
baseline=False,
epoch=None,
post=False,
multiple=True,
dataset='bam',
args=None):
res = []
for seed in seeds:
curr = []
if dataset == 'bam':
_, testloader = dataload.get_data_loader_SceneBAM(seed=seed,ratio=float(0.5), specific=specific)
elif dataset != 'coco':
_, testloader = dataload.get_data_loader_idenProf('idenprof',train_shuffle=True,
train_batch_size=64,
test_batch_size=64,
exclusive=True)
for ratio in ratios:
model, _, _, _ = load_models(
device,
base_folder=base_folder,
specific=specific,
seed=seed,
module=module,
experiment=experiment,
ratio=ratio,
adv=adv,
baseline=baseline,
epoch=epoch,
post=False, # n2v doesnt matter, we should at least have the during n2v trained
multiple=multiple,
leakage=False,
dataset=dataset,
args=args,
ignore_net=True
)
model.eval()
if dataset == 'coco':
tmp_args = copy.deepcopy(args)
tmp_args.ratio = 1
tmp_args.balanced = True
tmp_args.gender_balanced = True #False
if int(ratio) > 0:
tmp_args.balanced = True
_, testloader = coco_dataload.get_data_loader_coco(
tmp_args
)
with torch.no_grad():
acc = utils.detection_disparity(
testloader,
model,
device
)
else:
acc = utils.classification_specific_accuracy(
testloader,
model,
device
)
curr.append(acc)
res.append(curr)
return res
def load_models(device,
base_folder='./models/BAM/',
specific="bowling_alley",
seed=0,
module="layer3",
experiment="sgd_finetuned",
ratio="0.5",
adv=False,
baseline=False,
epoch=None,
post=False,
multiple=True,
leakage=False,
tcav=False,
force=False,
dataset='bam',
args=None,
ignore_net=False):
'''
if dataset == 'coco' and adv:
class DummyArgs:
num_object = 79
finetune=False
layer='generated_image'
autoencoder_finetune=True
finetune=True
model = balanced_models.ObjectMultiLabelAdv(DummyArgs(), 79, 300, True, 1)
ok = torch.load('model_best.pth.tar', encoding='bytes')
state_dict = {key.decode("utf-8"):ok[b'state_dict'][key] for key in ok[b'state_dict']}
model.load_state_dict(state_dict)
model.to(device)
model.eval()
'''
if leakage:
assert post
if epoch is not None:
epoch = "_" + str(epoch)
else:
epoch = ""
if len(args.custom_end) > 0:
args.custom_end = "_" + str(args.custom_end)
if baseline:
model_end = "resnet_base_"+str(ratio)+epoch+'.pt'
if not post:
n2v_end = "resnet_n2v_base_"+str(ratio)+epoch+'.pt'
else:
n2v_end = "resnet_n2v_base_after_"+str(ratio)+epoch+'.pt'
else:
if not adv:
model_end = "resnet_debias_"+str(ratio)+epoch+'.pt'
if not post:
n2v_end = "resnet_n2v_debias_"+str(ratio)+epoch+'.pt'
else:
n2v_end = "resnet_n2v_debias_after_"+str(ratio)+epoch+str(args.custom_end)+'.pt'
else:
model_end = "resnet_adv_"+str(ratio)+'.pt'
if not post:
n2v_end = "resnet_n2v_adv_"+str(ratio)+'.pt'
else:
n2v_end = "resnet_n2v_adv_after_"+str(ratio)+epoch+'.pt'
if dataset != 'bam' and dataset != 'coco':
model_end = model_end.replace('_'+str(ratio), '')
n2v_end = n2v_end.replace('_'+str(ratio), '')
if dataset == 'bam' or dataset == 'coco':
model_path, n2v_path = utils.get_paths(
base_folder,
seed,
specific,
model_end=model_end,
n2v_end='leakage/' + n2v_end.replace('n2v','mlp') if leakage else n2v_end,
n2v_module=module,
experiment=experiment,
with_n2v=True,
)
else:
model_path = os.path.join(base_folder, str(seed), experiment, module, model_end)
n2v_path = os.path.join(base_folder, str(seed), experiment, module, 'leakage/' + n2v_end.replace('n2v','mlp') if leakage else n2v_end)
if dataset == 'bam':
trainloader, _ = dataload.get_data_loader_SceneBAM(seed=seed,ratio=float(ratio), specific=specific)
_, testloader = dataload.get_data_loader_SceneBAM(seed=seed,ratio=float(0.5), specific=specific)
elif dataset == 'coco':
tmp_args = copy.deepcopy(args)
tmp_args.ratio = ratio
if int(ratio) > 0:
tmp_args.balanced = True
if leakage:
tmp_args.gender_balanced = True
trainloader, testloader = coco_dataload.get_data_loader_coco(
tmp_args
)
else:
trainloader,testloader = dataload.get_data_loader_idenProf('idenprof',train_shuffle=True,
train_batch_size=64,
test_batch_size=64,
exclusive=True)
if not (dataset == 'coco' and adv):
assert os.path.exists(model_path), model_path
if post:
# since we have to run a separate script, might not have finished...
if not leakage:
model_extra = '_adv' if adv else ('_base' if baseline else '_debias')
n2v_extra = model_extra + '_after'
if tcav:
pass
elif force:
post_train.train_net2vec(trainloader,
testloader,
device,
seed,
specific=specific,
p=ratio,
n_epochs=20,
module=module,
lr=.01,
out_file=None,
base_folder=base_folder,
experiment1=experiment,
experiment2=experiment,
model_extra=model_extra,
n2v_extra=n2v_extra,
with_n2v=True,
nonlinear=False, # might want to change this later
model_custom_end=epoch.replace('_',''),
n2v_custom_end=epoch.replace('_',''),
multiple=multiple,
dataset=dataset
)
else:
raise Exception('Run trial again')
elif leakage:
model_extra = '_adv' if adv else ('_base' if baseline else '_debias')
n2v_extra = model_extra + '_after'
if force:
post_train.train_leakage(trainloader,
testloader,
device,
seed,
specific=specific,
p=ratio,
n_epochs=20,
module=module,
lr=5e-5, # leakage model uses adam
out_file=None,
base_folder=base_folder,
experiment1=experiment,
experiment2=experiment,
model_extra=model_extra,
n2v_extra=n2v_extra,
with_n2v=True,
nonlinear=True, # MLP leakage model
model_custom_end='',
n2v_custom_end='',
dataset=dataset
)
else:
raise Exception('Run trial again')
else:
# should've been saved during training if not ported from tianlu
if not (dataset == 'coco' and adv):
assert os.path.exists(n2v_path)
num_attributes = 10 + 9 + 20 if multiple else 12
num_classes=10
if dataset == 'coco':
num_attributes = 81
num_classes = 79
model, net, net_forward, activation_probe = models.load_models(
device,
None if (dataset == 'coco' and adv) else
lambda x,y,z: models.resnet_(
pretrained=True,
custom_path=x,
device=y,
initialize=z,
num_classes=num_classes,
size=50 if (dataset == 'bam') or (dataset == 'coco') else 34
),
model_path=model_path,
net2vec_pretrained=True,
net2vec_path=n2v_path,
module='fc' if leakage else module,
num_attributes=2 if leakage else num_attributes,
model_init = False,
n2v_init = False,
nonlinear = leakage,
ignore_net = ignore_net
)
print(n2v_path)
return model, net, net_forward, activation_probe